Psychological consultation platform and method based on Multi-agent

Abstract

The invention discloses a Multi-agent-based psychological counseling platform and method. The counseling platform comprises an evaluation agent used for dynamically extracting psychological features based on dialogue behaviors of a visitor agent and updating results to a traits library module; the planning agent is used for constructing a personalized psychological counseling scheme based on the traits library module and dynamically adjusting a counseling process and a strategy according to counseling feedback; the consultant agent is used for carrying out dialogue interaction with the visitor agent according to the personalized consultation scheme and implementing emotion pacification and psychological intervention; the visitor agent is used for expressing psychological troubles, feeding back psychological state changes and promoting the consultation process; the historical dialogue long and short-term memory module is used for storing multiple rounds of psychological counseling interaction contents and comprises historical dialogue short-term memory and historical dialogue long-term memory; and the dynamic probability memory retrieval module is used for generating probability distribution based on the current consultation context and dynamically calling memory nodes from the historical dialogue long and short term memory module or the traits library module.

Description

Technical Field

[0001] The present application relates to the field of multi-agent large models, and in particular to a multi-agent-based psychological counseling platform and method. Background Art

[0002] In recent years, breakthroughs in artificial intelligence (AI) technology have provided a new paradigm for reshaping the mental health service system. Natural language processing (NLP), centered around large language models, is reshaping the delivery of psychological services. Researchers have leveraged NLP to explore topics including mental health assessment, enhancing emotional support, and mental health counseling. The development of large language models has not only advanced psychological counseling research but has also begun to be used as effective therapeutic aids. The latest trends in these models are generating more interpretable mental health analyses and simulating interactions between counselors and clients, providing a more authentic counseling experience.

[0003] Although the online psychological counseling systems or intelligent chatbots currently available on the market can alleviate users' emotions to a certain extent, they still have the following major shortcomings:

[0004] 1. Lack of personalization: Existing systems often rely on fixed processes or preset dialogue scripts, making it difficult to dynamically adjust to the unique psychological characteristics of each visitor. This often leads to monotonous consultation content and greatly reduced effectiveness.

[0005] 2. Memory is fragmented and gaps are obvious: Most platforms only retain short-term conversation records and lack the continuous accumulation and utilization of the long-term psychological characteristics of the visitors. During consultations, there are often situations where "the previous words are inconsistent" or important information discussed previously is forgotten.

[0006] 3. Single intervention strategy and delayed response: Existing intelligent interventions usually only provide general comfort or guidance after detecting negative emotions in users. They lack layered, continuous, and long-term psychological program planning, making it difficult to truly help users gradually improve their mental state.

[0007] 4. Imperfect evaluation and feedback mechanisms: Most platforms lack systematic effectiveness evaluation methods and rely solely on user satisfaction or simple sentiment analysis. They do not have a multi-dimensional evaluation mechanism that combines professional scales or manual reviews, and are unable to accurately quantify consulting quality and intelligent agent performance.

[0008] 5. Weak privacy and security protection: Some systems are not adequately encrypted or desensitized at the data storage and transmission levels, and lack detailed permission control and audit processes, posing a risk of user privacy leakage and undermining user trust and long-term use.

[0009] 6. Insufficient support for multimodal interaction: Currently, it is mainly based on text or voice, and there is little multi-source information fusion that combines expressions, facial expressions, heart rate and other physiological signals, resulting in an incomplete and inaccurate perception of the visitor's true emotions. Summary of the Invention

[0010] The purpose of the present invention is to overcome the deficiencies in the prior art and to provide a multi-agent-based psychological counseling platform and method.

[0011] The purpose of the present invention is achieved through the following technical solutions:

[0012] A multi-agent based psychological consultation platform, including:

[0013] The evaluation agent is used to dynamically extract psychological characteristics based on the conversational behavior of the visitor agent and update the extracted results to the trait library module;

[0014] Planning agent, used to build personalized psychological counseling plans based on the trait library module and dynamically adjust counseling processes and strategies based on counseling feedback;

[0015] The counselor agent is used to interact with the client agent based on the personalized counseling plan, and implement emotional comfort and psychological intervention;

[0016] The client agent is used to express psychological distress, provide feedback on changes in psychological state, and promote the consultation process;

[0017] The historical conversation long-term and short-term memory module is used to store the content of multiple rounds of psychological consultation interactions, including historical conversation short-term memory and historical conversation long-term memory;

[0018] The dynamic probabilistic memory retrieval module is used to generate a probability distribution based on the current consultation context and dynamically retrieve memory nodes from the historical dialogue long-term and short-term memory module or the trait library module.

[0019] Furthermore, the evaluation agent includes:

[0020] The trait library module consists of four sub-modules: basic information, intrinsic traits, extrinsic traits, and key events, and is used to dynamically build and maintain personalized psychological portraits;

[0021] The behavior assessment module is used to evaluate the visitor's psychological state and behavioral characteristics based on the current information and update the trait library module.

[0022] Furthermore, the planning agent adopts the dual-layer memory architecture to read and write information in the long-term and short-term memory modules of historical dialogues and the trait library module through the dual-layer memory architecture; the planning agent also includes:

[0023] The consultation planning module is used to develop a tiered consultation plan covering short-term emotional comfort and long-term psychological intervention;

[0024] Strategy adjustment module, used to dynamically optimize consulting solutions based on feedback from the current round;

[0025] Summarization behavior module, used to extract key content from the conversation and update long-term memory.

[0026] Furthermore, the consultant agent adopts a dual-layer memory architecture, which reads and writes information in the long-term and short-term memory modules of historical conversations and the trait library module to assist in consultation execution; the consultant agent also includes:

[0027] The dialogue behavior module is used to conduct consultation interactions with visitors based on the consultation plan provided by the planning agent;

[0028] Summarization behavior module, used to extract key content from the conversation and update long-term memory.

[0029] Furthermore, the visitor agent includes:

[0030] Personality module, used to store basic attributes, interpersonal communication attributes, role-playing attributes, and other psychological behavior attributes;

[0031] Memory module, used to store the client's perceptual memory, short-term memory and long-term memory based on personality type;

[0032] Speech behavior module, used to express psychological distress and summarize changes in psychological state;

[0033] Summarization behavior module, used to extract key content from the conversation and update long-term memory.

[0034] Furthermore, the trait library module is dynamically updated by the evaluation agent and can be called by the planning agent and consultant agent to assist in program design and strategy execution.

[0035] Furthermore, it also includes a multi-agent psychological counseling evaluation system, which combines automatic evaluation indicators with manual evaluation indicators to systematically measure the counseling effect and agent performance.

[0036] Furthermore, the recency score calculation method and dynamic probabilistic memory extraction mechanism are improved in the dynamic probabilistic memory retrieval module to optimize the selection of memory nodes;

[0037] Specifically, a ranked recency score is used to ensure that the most recent memory nodes are given higher weights and the most distant memory nodes are given lower weights, in order to strengthen the priority of short-term memory in the retrieval process;

[0038] A dynamic probabilistic memory extraction mechanism is adopted to normalize the comprehensive score into a probability distribution, and k memory nodes are randomly selected based on the distribution.

[0039] Furthermore, the dynamic probabilistic memory retrieval module improves the classic retrieval scoring function, comprehensively considering the recency score, relevance score, and importance score, including:

[0040] Recency score: This is calculated based on time sorting, with the oldest memory receiving the lowest score and the most recent memory receiving the highest score. Specifically, all memory objects are sorted from earliest to latest by creation time, with the oldest memory ranked 1 and the most recent memory ranked N, where N is the total number of memories. The ranking score is mapped to the [0,1] interval using the Min-Max normalization method.

[0041] Relevance scoring: Based on semantic space similarity calculation, a memory association network is constructed. Specifically, the current query content and the text description of the trait library module are converted into embedding vectors, the cosine similarity is calculated, and the similarity value range is mapped to the [0,1] interval using the Min-Max normalization method.

[0042] Importance Scoring: Semantic analysis is used to distinguish between general and key trait information, and a large language model (LLM) direct scoring mechanism is used. Specifically, the LLM is used to quantitatively evaluate the importance of trait information on a scale of 1-10, and the Min-Max normalization method is used to map it to the interval [0, 1]. This score is calculated when the object in the trait library module is created.

[0043] The final retrieval score is obtained by calculating the above three scores using a linear weighted fusion method.

[0044] The present invention also provides a psychological consultation method, based on the above-mentioned psychological consultation platform, comprising:

[0045] Step 1: Initialize the visitor agent and have it initiate a consultation request;

[0046] Step 2: The evaluation agent analyzes and updates the information in the visitor's psychological state and trait library module;

[0047] Step 3: The planning agent formulates or optimizes a personalized consulting plan based on the trait library module;

[0048] Step 4: The counselor agent implements personalized psychological intervention dialogue;

[0049] Step 5: The client's intelligent agent reflects the psychological changes and promotes the continuation of the consultation round;

[0050] Repeat steps 2 to 5 until the consultation goal is achieved or the conversation ends.

[0051] Compared with the prior art, the technical solution of the present invention has the following beneficial effects:

[0052] 1. Enhanced Personalization and Dynamic Response: By introducing planning and evaluation agents, this invention builds a comprehensive multi-agent collaborative system, enabling real-time perception of the client's psychological state, feature modeling, and dynamic adjustment of consultation strategies. This system flexibly tailors consultation paths based on the user's current emotional state and historical interaction information, ensuring each conversation is closely aligned with the user's needs, avoiding "one-size-fits-all" interventions and significantly improving the adaptability and individualization of the consultation process.

[0053] 2. High-fidelity Psychological Trait Modeling Mechanism: Through an innovative dual-layer memory architecture and dynamic probabilistic memory retrieval module, this invention effectively addresses the issues of memory gaps and knowledge forgetting during the psychological trait modeling process. The system not only maintains a coherent understanding of context but also accumulates user psychological characteristics over time, enabling deeper emotional analysis and behavioral prediction, providing a reliable basis for personalized psychological intervention.

[0054] 3. Adaptive optimization and continuity assurance of the consultation process: The planning agent has the ability to autonomously adjust the consultation strategy and optimize the direction of the dialogue according to the client's feedback in multiple rounds of interaction, ensuring the consistency and depth of intervention in the psychological consultation process, and effectively improving the counseling efficiency and client satisfaction.

[0055] 4. A systematic evaluation system promotes the continuous evolution of intelligent agents: This paper constructs an intelligent evaluation system that integrates automatic assessment scale score changes (such as changes in sentiment scores) with manual assessments (such as credibility, professionalism, engagement, and safety). This system can comprehensively and objectively measure the effectiveness of psychological counseling and system performance. The evaluation results not only support counseling quality control but also provide a scientific basis for the continuous optimization and parameter adjustment of the intelligent agent model, promoting the continuous improvement of the system's intelligence level.

[0056] 5. Supports multiple scenarios and promotes professional training: This platform is applicable to a variety of scenarios, including individual counseling, simulation training, and scientific research analysis. It is particularly valuable in the practical training and competency assessment of psychological counselors. The system can simulate various individual psychological states and coping behaviors, providing a realistic and accurate training environment for psychological counselors, helping them improve their intervention capabilities and professional level.

[0057] 6. Promote the integration of psychological research and technology: By simulating the evolution of individual behavior and psychological changes, the platform provides authentic and reliable data support for psychology researchers, helping to expand the empirical basis of psychological counseling theories. Furthermore, this invention promotes the deep integration of artificial intelligence and psychological counseling technology, opening up innovative paths in the field of mental health services.

[0058] In summary, the present invention has made significant progress in the intelligence, systematization and personalization of psychological counseling, overcomes many bottleneck problems in the existing technology, and has good scientific research value, application prospects and promotion potential. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 This is the overall flow chart of the psychological consultation platform implemented based on Multi-agent in the present invention; DETAILED DESCRIPTION

[0060] The following will be combined with the accompanying drawings in the embodiments of the invention to clearly and completely describe the technical solutions in the embodiments of the invention. Obviously, the embodiments described are only part of the embodiments of this application, not all of them. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0061] This embodiment provides a multi-agent-based psychological counseling platform, including an assessment agent, a planning agent, a counselor agent, a visitor agent, a historical dialogue long-term and short-term memory module, and a dynamic probabilistic memory retrieval module; the psychological counseling platform improves the consistency, personalization, and intervention effect of the counseling process through clear division of labor and collaborative work.

[0062] The assessment agent is responsible for continuously and dynamically assessing the client's psychological state and continuously updating the trait library module with the assessment results. The trait library module records the client's basic information, psychological characteristics, emotional changes, and key psychological events, providing personalized support for the consultation process. Through this mechanism, the client's psychological state can be continuously tracked, the long-term evolution of psychological traits can be captured, and a complete and stable personalized psychological profile can be constructed, improving the ability to identify deep psychological needs.

[0063] Based on the trait memory information accumulated by the evaluation agent, the planning agent develops personalized counseling plans that balance short-term emotional soothing and long-term psychological intervention. It optimizes counseling plans based on real-time feedback during the consultation process, dynamically adjusts response strategies for the current round, and provides conversational guidance for the counselor agent. This module addresses the shortcomings of traditional systems in dynamically planning multi-stage counseling goals, enabling the counseling platform to accurately match individual needs and achieve an effective balance between short-term emotional soothing and long-term psychological intervention, thereby improving the consistency and effectiveness of counseling across rounds.

[0064] The counselor agent relies on the trait library and psychological counseling plan of the evaluation agent and planning agent to carry out specific psychological counseling tasks and conduct psychological counseling dialogues with visitors.

[0065] The client agent simulates the status and needs of real psychological help seekers and expresses psychological distress through interaction with the counselor agent.

[0066] The historical conversation long-term and short-term memory module is used to store the content of multiple rounds of psychological consultation interactions, including historical conversation short-term memory and historical conversation long-term memory;

[0067] The dynamic probabilistic memory retrieval module is used to generate probability distribution based on the current consultation context and dynamically retrieve memory nodes from the historical dialogue long-term and short-term memory module or the trait library module.

[0068] In this embodiment, the role set in the psychological counseling platform is defined as Where C represents a set of roles, c i Indicates a specific role.

[0069] Specifically:

[0070] (1) Evaluating the Agent

[0071] During its interactions with visitors, the assessment agent continuously analyzes and updates its trait library module to construct a more accurate and dynamic visitor profile. This agent's assessment covers multiple dimensions, including basic information, intrinsic traits, extrinsic traits, and key events, ensuring a comprehensive understanding of the visitor's personality traits, psychological state, and behavioral patterns. The assessment agent consists of an assessment behavior module and a trait library module. The trait library module is updated after perceiving the visitor's speech and is used to accumulate and maintain the visitor's personality traits, dynamically improving the visitor's profile. The assessment behavior module is used to integrate current information to assess the visitor's psychological state and behavioral characteristics and update the trait library module.

[0072] The trait library module is composed as follows:

[0073] Basic information submodule: records the visitor's objective background information, including age, gender, educational background, occupation, interests and hobbies, etc., to provide a basic reference for subsequent evaluation.

[0074] Intrinsic traits sub-module: focuses on the client's psychological and cognitive characteristics, including emotional state, cognitive patterns, psychological needs, personality type, and psychological defense mechanisms, in order to analyze their thinking style, emotional regulation ability, and deep psychological needs.

[0075] External trait submodule: Evaluate the client's behavioral performance, interpersonal relationships, social support system, living conditions and physical health status, and improve their trait portrait from the perspective of the interaction between external behavior and the environment.

[0076] Key Events Sub-Module: Record important experiences that affect the client, including developmental events in growth, stressful events in life and their impact, and analyze the process of shaping their psychology and behavior.

[0077] (2) Planning agent, including consulting planning module, strategy adjustment module and behavior summary module.

[0078] The planning agent formulates a personalized psychological counseling plan based on the visitor's trait information. By designing a hierarchical psychological counseling model, short-term emotional soothing and long-term psychological intervention plans are planned, and the psychological counseling strategies adopted in the current round are analyzed. For the planning agent, this embodiment designs a two-layer memory architecture based on the long-term and short-term memory modules and the trait library modules of historical dialogues to simulate the counselor's thinking mode in the psychological counseling planning process. During the consultation dialogue, the planning agent can continuously optimize the personalized hierarchical psychological counseling plan based on the dynamically updated user trait portrait. The behaviors of the planning agent include planning, speaking and summarizing. Among them, the planning behavior is used to formulate or update the psychological counseling plan, and the summarizing behavior is responsible for extracting key information in the dialogue history and updating it to the long-term memory to support subsequent decision-making and strategy adjustments. The memory structure is a two-layer memory architecture.

[0079] (3) Consultant agent, including dialogue behavior module and summary behavior module.

[0080] The counselor agent provides personalized psychological counseling to the visitor based on the information in the visitor's trait library module and the counseling plan and strategy provided by the planning agent. For the counselor agent, this embodiment designs a two-layer memory architecture based on the long-term and short-term memory module and the trait library module of the historical dialogue to simulate the counselor's thinking mode during the psychological counseling process. During the consultation dialogue, the counselor agent implements targeted psychological intervention and dialogue support based on the dynamically updated visitor's trait portrait and personalized layered psychological counseling plan. The behavior of the counselor agent includes speaking and summarizing. Among them, the summary behavior is used to refine and summarize the key information in the dialogue history and store it in the long-term memory to provide information support for subsequent consultations. The memory structure is a two-layer memory architecture.

[0081] (4) Visitor intelligent body, including personality module, memory module, speech behavior module and summary behavior module.

[0082] The personality module consists of four parts:

[0083] Basic attributes: name, gender, age, personality type, educational background, occupation, personality, interests and specialties, interpersonal relationships, growth experience, motivation for participation, whether there is psychological counseling experience, and willingness to participate in psychological counseling.

[0084] Interpersonal communication attributes: self-cognition, interpersonal attitude, and interpersonal perception.

[0085] Role-playing attributes: language characteristics, emotional expression, and interaction patterns.

[0086] Other attributes: extroversion index, problems faced.

[0087] The memory module includes perceptual memory, short-term memory, and personality-based long-term memory. Perceptual memory stores the perceived speech content of the visitor agent in the current environment. The visitor agent's behavior includes speaking and summarizing. The speaking behavior module expresses psychological distress and summarizes changes in mental state. The summarizing behavior module extracts and summarizes the conversation history based on the agent's personality type and long-term memory preferences, and stores key information in long-term memory to support subsequent conversations and psychological state expression.

[0088] Preferably, this embodiment also provides a psychological counseling method based on the above psychological counseling platform, see Figure 1 The specific process is as follows:

[0089] Step 1: Initialize the client agent. The client agent simulates the role of a real psychological help seeker, actively initiates a consultation request and expresses psychological distress, marking the official start of the psychological counseling process.

[0090] Step 2: The assessment agent conducts a comprehensive analysis of the client's psychological state and individual traits, and updates the assessment results to the trait library module in real time. The trait library module primarily stores the client's basic information, psychological characteristics, emotional changes, and key psychological events, providing precise and personalized support for subsequent consultations.

[0091] Step 3: Based on the information in the trait library module, the planning agent designs or updates personalized psychological counseling plans for short-term emotional comfort and long-term psychological intervention, and determines the counseling strategy for the current round to provide dynamic guidance for the counseling process.

[0092] Step 4: The counselor agent comprehensively refers to the client's characteristics and counseling plan, conducts targeted psychological counseling dialogues with the client's agent, and implements specific measures such as emotional support, cognitive regulation, and psychological intervention.

[0093] Step 5: Client Agent Feedback and Conversation Advancement: The client agent provides feedback on its psychological state and changes in needs based on the current consultation content, further advancing the consultation process to the next round of interaction. The system continues to loop through Steps 2 to 5 until the consultation goal is achieved or the conversation ends.

[0094] Preferably, the psychological counseling platform provided in this embodiment involves a dual-layer memory architecture of a long-term and short-term memory module based on historical dialogues and a trait library module, which is used to simulate the thinking and decision-making process of a real psychological counselor and fully support information management and strategy planning in psychological counseling. This architecture can dynamically construct a visitor portrait during the consultation process, continuously track changes in psychological state, plan short-term emotional comfort and long-term psychological intervention plans, and implement personalized psychological counseling. Specifically, the dual-layer memory architecture includes the following two parts: (1) Historical dialogue long-term and short-term memory module: used to store and manage short-term memory and long-term memory in multiple rounds of dialogue. (2) Trait library module: dynamically maintained by the evaluation agent, used to record individual traits such as the visitor's basic information, internal traits, external traits, and key psychological events, providing long-term, stable, and personalized support for psychological counseling program design and intervention strategy selection. The planning agent and the counselor agent can retrieve trait information related to the current round of responses in the trait library module to assist in decision-making and execution. Based on a dual-layer memory architecture, the planning agent can comprehensively call upon historical conversations and long-term trait information, dynamically formulate and optimize psychological counseling plans, and balance short-term emotional soothing with long-term psychological intervention goals; the counselor agent relies on the support of dual-layer memory to conduct coherent and accurate psychological counseling conversations.

[0095] Long-term and short-term memory based on historical conversations: This memory structure includes perceptual memory, short-term memory, and long-term memory. Perceptual memory directly interacts with the environment, that is, it imitates human perception of the environment and converts the perceived external observations into message objects and stores them in short-term memory. The output of perceptual memory is the message object M=<c,n,t,s> , representing the specific content of the action, the name of the person performing the action, the time of the action, and the importance rating of the action, respectively. The short-term memory module records the character's behavioral history during group psychological counseling, forming an irreversible timeline. Only new content is allowed to be added, not deleted or modified, thus ensuring the historical consistency of the character's behavior. Short-term memory is the intermediate link between perceptual memory and long-term memory. It records the agent's behavior and observable behavior. If the agent frequently encounters similar observations, short-term memory is updated to long-term memory. According to memory style theory and cognitive style theory, individuals exhibit stable personality differences in the encoding, storage, and retrieval of information. These differences directly influence the selection and expression of memory content. Based on this, a personality-driven long-term memory aggregation mechanism is proposed. By introducing personality type to dynamically regulate the generation and retrieval of long-term memory, each agent can maintain personalized long-term memory based on its own memory preferences. The personality type-based long-term memory aggregation mechanism combines the long-term memory aggregation mechanism with the personality type-based long-term memory regulation mechanism. While ensuring efficient long-term memory retrieval, it also aligns memory content with the character's personality, thereby strengthening the consistent expression of the character. The personality-type-based memory regulation mechanism aims to ensure personalized expression of long-term memory. Characters with different personality types exhibit significant differences in memory preferences. This mechanism leverages the cue word capabilities of the large language model to guide information summarization and memory updating based on the character's personality type after each conversation. It dynamically determines whether to retain key information or deemphasize secondary content, ensuring that the memory style matches the character's characteristics. This approach enables the agent to maintain a consistent character style throughout long-term conversations, enhancing the coherence and authenticity of the conversations. Throughout group counseling, each character maintains independent long-term memory, which serves as a global summary of past conversations. Long-term memory updates follow the following process: Short-term memory from the current turn is refined to form a short-term memory summary. Based on the character's personality type, the personality-type-based long-term memory regulation mechanism determines reinforcement and filtering rules. The long-term memory aggregation mechanism then merges the filtered short-term memory summary with the historical long-term memory to generate a new long-term memory. This process is repeated to continuously aggregate and optimize long-term memory.

[0096] Dynamic probabilistic memory retrieval module: This embodiment proposes an improved recency score calculation method and a dynamic probabilistic memory extraction mechanism in the dynamic probabilistic memory retrieval module to optimize the selection of memory nodes. Specifically, a sorted recency score is adopted to ensure that the most recent memory node obtains a higher weight, while the farthest memory node has a lower weight, thereby strengthening the priority of recent memory in the retrieval process. A dynamic probabilistic memory extraction mechanism is adopted to normalize the comprehensive score into a probability distribution, and k memory nodes are randomly selected based on the distribution. The advantage of this strategy is that high-frequency knowledge still has a greater probability of being selected, but it will not completely monopolize the retrieval results; although low-frequency knowledge has a lower probability of being selected, it still has the opportunity to enter the retrieval range, avoiding valuable information from being forgotten due to low access frequency, thereby improving the flexibility and information coverage of the system. The classic retrieval scoring function is improved, and the recency score, relevance score and importance score are comprehensively considered. The specific design is as follows:

[0097] Recency score N: Calculated based on time order, with the oldest memory receiving the lowest score and the most recent memory receiving the highest score. The specific process is as follows: All memories are sorted from earliest to latest creation time, with the oldest memory ranked 1 and the most recent memory ranked N (of the total number of memories). Min-Max normalization is used to map the ranking scores to the [0, 1] interval.

[0098] Relevance score R: Based on semantic space similarity calculation, a memory association network is constructed. The specific process is: the current query content and the text description of the trait library module are converted into embedding vectors, cosine similarity is calculated, and the similarity value range is mapped to the [0, 1] interval using the Min-Max normalization method.

[0099] Importance Scoring I: This method uses semantic analysis to distinguish between common and key traits, employing a direct scoring mechanism using the Large Language Model (LLM). In its implementation, the LLM is prompted to quantify the importance of traits on a scale of 1-10, mapping them to the [0, 1] interval using a Min-Max normalization method. This score is calculated when objects are created in the trait library module.

[0100] The final retrieval score adopts a linear weighted fusion strategy:

[0101] score=α n N+α r R+α i I

[0102] In the experimental setting, the weight coefficient α n ,α r ,α i Both are 1.

[0103] High-frequency knowledge, due to its high relevance and importance, will occupy the top position for a long time, resulting in a solidification phenomenon. Rare knowledge, even if occasionally useful, cannot enter the top-k position, causing the system to forget it. Therefore, this module introduces a probabilistic sampling strategy to normalize the comprehensive score into a probability distribution:

[0104]

[0105] We randomly select k memory nodes (k = 10 in the experiment) based on a probability distribution to improve the system's coverage of long-tail knowledge. This module's design balances the retrieval opportunities of high-frequency knowledge with low-frequency, effective knowledge through a probabilistic sampling mechanism. Compared to a fixed Top-k approach, this significantly improves the diversity and adaptability of psychological counseling in counseling scenarios.

[0106] This embodiment uses two types of evaluation methods to evaluate the performance of the large language model (LLM), namely automatic evaluation and manual evaluation indicators.

[0107] Before the psychological consultation begins, the client is tested on the psychological scale, and the test result is scale_pre. After the psychological consultation ends, the client is tested on the psychological scale, and the test result is scale_after. The scale change is s_change.

[0108] The emotional score of the first speech of the client in psychological consultation is emotion_pre, the emotional score of the last speech of the client in psychological consultation is emotion_after, and the change in the emotional score is e_change.

[0109] The total output length of the agent is length, the number of times the agent speaks is number, and the average output length is averge_length. Specifically, calculate the output length and number of times the consultant agent speaks and the client agent speaks separately.

[0110] Scale Change: Across multiple experiments, the average change in the client's scale test scores before and after counseling was calculated. The greater the scale change, the better the Large Language Model (LLM) was. The scale used was the Positive Mental Health Scale (PMHS).

[0111]

[0112] Sentiment score change: Across multiple experiments, the average change in the counseling client's sentiment scores was calculated. The greater the change in sentiment score, the better the Large Language Model (LLM) performance. The sentiment score was calculated using the HanLP sentiment analysis model.

[0113]

[0114] Average output length; calculate the average output length of counselors and clients separately.

[0115] leader_average_length=main_length / main_number

[0116] seeker_average_length=seeker_length / seeker_number

[0117] This example also proposes a four-dimensional manual evaluation index system to comprehensively evaluate the agent-simulated psychological counseling based on credibility, professionalism, engagement, and safety. The evaluation is performed by the GPT-4 model, which independently scores each dimension within a fixed range [lower bound – upper bound]. The specific definitions and scoring criteria for each dimension are detailed in Table 1.

[0118] Credibility [0-5]: This evaluates the agent's authenticity in the simulated counseling session, specifically the plausibility and naturalness of its behavior in the dialogue between the client and counselor. A higher score indicates a more realistic representation of the interaction pattern in a real counseling session.

[0119] Professionalism [0-6]: Assess the counselor agent's professionalism in counseling, focusing on their overall performance in questioning techniques, emotional support, and relationship building during the counseling process. Higher scores indicate stronger counseling expertise.

[0120] Participation [0-3]: This assesses the client's level of active expression during the consultation process, including their presentation of their own problems, emotional expression, and feedback on their needs. A higher score indicates a higher degree of client participation in the conversation and a more thorough presentation of their problems.

[0121] Safety [0-2]: Assessing the correctness of values ​​and ethical safety during the consultation process, including whether the ethical standards of psychological counseling are followed and whether the client's privacy and sensitive information are protected. The higher the score, the more ethically sound and safe the consultation process is.

[0122] Table 1: Detailed evaluation criteria for manual evaluation indicators

[0123]

[0124] Psychological counseling system: GPT-4 and Qwen-Max are used as the backbone LLMs of the role agents in the psychological counseling platform. The Qwen-Max model is used in the Chinese context, and the GPT-4 model is used in the English context. In addition, two ablation systems are implemented to study the effects of different modules in the psychological counseling platform. In the first system (DLMA-ICF-w / oplanleader), the planning agent is removed. In the second system (DLMA-ICF-w / oassessleader), the assessment agent and trait library modules are removed.

[0125] Counseling Participants: In the experiment, one planning agent, one evaluator agent, one counselor agent, and one client agent were used to conduct the counseling. Each combination of LLM settings and ablation system was repeated five times.

[0126] To systematically evaluate the effectiveness of the proposed psychological counseling platform, DLMA-ICF, this example selected representative baseline models for comparison, including ChatGPT, Qwen-Max, SoulChat, CoE, and AutoCBT, based on four dimensions: general large-scale model, mental health large-scale model, prompt engineering method, and multi-agent system design. The platform's psychological counseling evaluation metrics were compared with those of the baseline models. The automated evaluation results of the comparative experiment are shown in Table 2, and the manual evaluation results are shown in Table 3.

[0127] Table 2 shows the automatic evaluation results of different models in Chinese and English psychological counseling tasks. The DLMA-ICF platform achieved the best performance in indicators such as scale change, sentiment score change, and dialogue output length. The scale change of the DLMA-ICF platform in Chinese and English scenarios reached 2.28 and 2.16 respectively, which is significantly higher than other baseline models, indicating that it can more effectively improve the psychological state of the client; the sentiment score change is also the highest, indicating that it has a more obvious effect on the client's emotional regulation; the generated dialogue content is richer, and the average output length of both the counselor and the client exceeds that of other models. In comparison, AutoCBT's overall performance is second best, and MindChat, CoE, and Qwen-Max (GPT-4) are relatively weak. The DLMA-ICF platform performed outstandingly in the psychological counseling dialogue generation task.

[0128] Table 3 shows the manual evaluation results of different models in Chinese and English psychological counseling tasks. The DLMA-ICF model achieved the highest scores across all indicators in both languages, achieving the best overall performance. DLMA-ICF achieved credibility, professionalism, engagement, and security scores of 4.54, 5.62, 2.73, and 1.98, respectively, in the Chinese scenario; and 4.55, 5.65, 2.68, and 1.97, respectively, in the English scenario, all outperforming other baseline models. DLMA-ICF not only provides more credible and professional psychological counseling services, but also effectively improves conversational interactivity and ensures a high level of security. DLMA-ICF demonstrated consistent advantages in manual evaluations, validating its effectiveness and reliability in psychological counseling dialogue generation tasks.

[0129] Table 2: Automatic evaluation results of comparative experiments

[0130]

[0131] Table 3: Manual evaluation results of comparative experiments

[0132]

[0133]

[0134] To verify the effectiveness of each component, we conducted ablation experiments on the Large Language Model (LLM), removing different modules and evaluating the performance changes. The ablation results are shown in the table. "-w / o assessleader" indicates the removal of the assessment agent and trait library modules, and "-w / o planleader" indicates the removal of the planning agent module. The automatic evaluation results of the ablation experiments are shown in Table 4, and the manual evaluation results are shown in Table 5.

[0135] According to the automated evaluation results of the ablation experiment, the evaluation metrics of the psychological counseling platform decreased to varying degrees after removing the corresponding modules, thus verifying the effectiveness of each module. Specifically, after removing the planning agent module, the scale change, the emotion score change index, and the average output length of both counselors and clients all decreased significantly, confirming the effectiveness of this module setup and its ability to improve psychological counseling outcomes. After removing the evaluation agent and trait library modules, the scale change, the emotion score change index, and the average output length of both counselors and clients all decreased significantly, confirming the effectiveness of this module setup and its ability to improve psychological counseling outcomes. The average output length of psychological counseling sessions in the Chinese context was longer than in the English context, a result consistent with the group counseling experiments in Chapter 3. After removing the planning agent, the average output length of clients decreased significantly, indicating that this module played a role in encouraging user input. After removing the evaluation agent, the average output length of clients also decreased, indicating that this module played a role in encouraging user input.

[0136] According to manual evaluation results from ablation experiments, removing the corresponding modules resulted in varying degrees of decline in the psychological counseling platform's credibility, professionalism, engagement, and security. The platform effectively simulates psychological counseling, effectively addresses client problems, and enhances counseling effectiveness. The experimental results validate the effectiveness of the designed platform's planning agent module, evaluation agent module, and trait library module.

[0137] Table 4: Automatic evaluation results of ablation experiments

[0138]

[0139] Table 5: Manual evaluation results of ablation experiments

[0140]

[0141] The present invention is not limited to the embodiments described above. The above description of the specific embodiments is intended to describe and illustrate the technical solutions of the present invention. The above specific embodiments are merely illustrative and not restrictive. Without departing from the scope of the present invention and the scope of protection of the claims, those skilled in the art may make various specific modifications based on the teachings of the present invention, all of which fall within the scope of protection of the present invention.

Claims

1. A multi-agent-based psychological consultation platform, characterized by: include; The evaluation agent is used to dynamically extract psychological characteristics based on the conversational behavior of the visitor agent and update the extracted results to the trait library module; Planning agent, used to build personalized psychological counseling plans based on the trait library module and dynamically adjust counseling processes and strategies based on counseling feedback; The counselor agent is used to interact with the client agent based on the personalized counseling plan, and implement emotional comfort and psychological intervention; The client agent is used to express psychological distress, provide feedback on changes in psychological state, and promote the consultation process; The historical conversation long-term and short-term memory module is used to store the content of multiple rounds of psychological consultation interactions, including historical conversation short-term memory and historical conversation long-term memory; The dynamic probabilistic memory retrieval module is used to generate a probability distribution based on the current consultation context and dynamically retrieve memory nodes from the historical dialogue long-term and short-term memory module or the trait library module.

2. The psychological counseling platform according to claim 1, characterized in that: The evaluation agent includes: The trait library module consists of four sub-modules: basic information, intrinsic traits, extrinsic traits, and key events, and is used to dynamically build and maintain personalized psychological portraits; The behavior assessment module is used to evaluate the visitor's psychological state and behavioral characteristics based on the current information and update the trait library module.

3. The psychological counseling platform according to claim 1, characterized in that: The planning agent adopts the dual-layer memory architecture to read and write information in the long-term and short-term memory modules of historical dialogues and the trait library module; The planning agent also includes: The consultation planning module is used to develop a tiered consultation plan covering short-term emotional comfort and long-term psychological intervention; Strategy adjustment module, used to dynamically optimize consulting solutions based on feedback from the current round; Summarization behavior module, used to extract key content from the conversation and update long-term memory.

4. The psychological counseling platform according to claim 1, characterized in that: The consultant agent adopts a dual-layer memory architecture, which reads and writes information in the long-term and short-term memory modules of historical conversations and the trait library module to assist in consultation execution; the consultant agent also includes: The dialogue behavior module is used to conduct consultation interactions with visitors based on the consultation plan provided by the planning agent; Summarization behavior module, used to extract key content from the conversation and update long-term memory.

5. The psychological consultation platform according to claim 1, characterized in that: The visitor agent includes: Personality module, used to store basic attributes, interpersonal communication attributes, role-playing attributes, and other psychological behavior attributes; Memory module, used to store the client's perceptual memory, short-term memory and long-term memory based on personality type; Speech behavior module, used to express psychological distress and summarize changes in psychological state; Summarization behavior module, used to extract key content from the conversation and update long-term memory.

6. The psychological counseling platform according to claim 1, characterized in that: The trait library module is dynamically updated by the evaluation agent and is called by the planning agent and consultant agent to assist in program design and strategy execution.

7. The psychological consultation platform according to any one of claims 1 to 6, characterized in that: It also includes a multi-agent psychological counseling evaluation system, which combines automatic evaluation indicators with manual evaluation indicators to systematically measure counseling effectiveness and agent performance.

8. The psychological consultation platform according to any one of claims 1 to 6, characterized in that: The dynamic probabilistic memory retrieval module has improved the recency score calculation method and dynamic probabilistic memory extraction mechanism to optimize the selection of memory nodes; Specifically, a ranked recency score is used to ensure that the most recent memory nodes are given higher weights and the most distant memory nodes are given lower weights, in order to strengthen the priority of short-term memory in the retrieval process; A dynamic probabilistic memory extraction mechanism is adopted to normalize the comprehensive score into a probability distribution, and k memory nodes are randomly selected according to the distribution.

9. The psychological consultation platform according to any one of claim 8, characterized in that: The dynamic probabilistic memory retrieval module improves the classic retrieval scoring function, comprehensively considering the recency score, relevance score, and importance score. Specifically, it includes: Recency score: This is calculated based on time sorting, with the oldest memory receiving the lowest score and the most recent memory receiving the highest score. Specifically, all memory objects are sorted from earliest to latest by creation time, with the oldest memory ranked 1 and the most recent memory ranked N, where N is the total number of memories. The ranking score is mapped to the [0,1] interval using the Min-Max normalization method. Relevance scoring: Based on semantic space similarity calculation, a memory association network is constructed. Specifically, the current query content and the text description of the trait library module are converted into embedding vectors, the cosine similarity is calculated, and the similarity value range is mapped to the [0,1] interval using the Min-Max normalization method. Importance Scoring: Semantic analysis is used to distinguish between general and key trait information, and a large language model (LLM) direct scoring mechanism is used. Specifically, the LLM is used to quantitatively evaluate the importance of trait information on a scale of 1-10, and the Min-Max normalization method is used to map it to the interval [0, 1]. This score is calculated when the object in the trait library module is created. The final retrieval score is obtained by using the linear weighted fusion method to calculate the above three scores.

10. A psychological counseling method, based on the psychological counseling platform according to any one of claims 1 to 9, characterized in that: include: Step 1: Initialize the visitor agent and have it initiate a consultation request; Step 2: The evaluation agent analyzes and updates the information in the visitor's psychological state and trait library module; Step 3: The planning agent formulates or optimizes a personalized consulting plan based on the trait library module; Step 4: The counselor agent implements personalized psychological intervention dialogue; Step 5: The client's intelligent agent reflects the psychological changes and promotes the continuation of the consultation round; Repeat steps 2 to 5 until the consultation goal is achieved or the conversation ends.

Images